Issues

• [Bug]: v1.32.1 tag has version 1.33.0

  cat-bug
  NVIDIA/cudnn-frontend#2205 · Filter by author tiran opened on Nov 2, 2026
  1

• [Bug]: SDPA backward (backend engines 5 / 7) gives wrong dQ / dK, NaN dQ or an illegal memory access when O, dO or dQ spans more than 4 GiB (cuDNN 9.27.0.42; GH200 and B300)

  NVIDIA/cudnn-frontend#2174 · Filter by author 0z5a opened on Nov 1, 2026

• SM107 paged D192/V128 prefill: default dispatch lands on backend engine 8, which faults (illegal address) past 2^31-byte Q

  area:global_attentioncat-bugmod-backendorig-nv-eng
  NVIDIA/cudnn-frontend#2150 · Filter by author YangXu1990uiuc opened on Nov 1, 2026

• MXFP8 SDPA backward: dK_amax came back NaN on a deterministic rerun (once on GB200 CI)

  NVIDIA/cudnn-frontend#2122 · Filter by author YangXu1990uiuc opened on Oct 31, 2026
  1

• python/*/cutedsl:be-9.27.0-cutedsl-4.8.1:sm100 7 jobs failed at test_bsa_fp8_blk128_dispatch_preserves_v_layout_and_metadata

  cat-bugmod-frontendorig-qa
  Bug

  NVIDIA/cudnn-frontend#2098 · Filter by author liujane-dev opened on Oct 31, 2026

• python/gemm/frost:be-9.27.0-cutedsl-4.8.0:sm100 3 jobs failed at test_sm120_quant_scale_k_padding_is_zeroed

  cat-bugmod-frontendorig-qa
  Bug

  NVIDIA/cudnn-frontend#2097 · Filter by author liujane-dev opened on Oct 31, 2026

• [Bug]: MOE_GROUPED_MATMUL with MXFP8 block scales reads the wrong token-scale rows for every group after the first (B200, cuDNN 9.26.0.51, CUDA 13.4 libraries); same cuDNN under CUDA 13.0 is correct

  NVIDIA/cudnn-frontend#2081 · Filter by author Theodore-Liu opened on Oct 30, 2026
  1

• [Feature]: Exposing Indexer, Compressor, and DSA Kernels to JAX

  cat-feature
  NVIDIA/cudnn-frontend#2065 · Filter by author jberchtold-nvidia opened on Oct 30, 2026

• SDPA bwd THD on SM80: wrong gradients when dO has its own token stride / ragged offset (backend eng7)

  area:global_attention
  NVIDIA/cudnn-frontend#2051 · Filter by author YangXu1990uiuc opened on Oct 29, 2026

• [Bug]: HSTU q_stage=1 reuses the epilogue barrier across independent warp groups

  NVIDIA/cudnn-frontend#2002 · Filter by author JacoCheung opened on Oct 29, 2026
  

• SM80 FROST SDPA rows stage gather/scatter copies on layouts they now serve by default (Rule 2)

  area:frost
  NVIDIA/cudnn-frontend#1986 · Filter by author YangXu1990uiuc opened on Oct 28, 2026
  1

• [Perf]: backend FP8 d128 SDPA fwd on Rubin, bottom-right causal with short s_q does not scale with heads/KV split (up to 12x slower than the mask-free graph)

  cat-perf-bugmod-backendorig-nv-eng
  NVIDIA/cudnn-frontend#1788 · Filter by author YangXu1990uiuc opened on Oct 22, 2026

• SM100 FP8 SDPA forward under-reports Amax_S at short s_q (max over the first 128 KV columns only)

  area:global_attentioncat-bugmod-backendorig-nv-eng
  NVIDIA/cudnn-frontend#1613 · Filter by author YangXu1990uiuc opened on Oct 13, 2026

• SM100 SDPA backward d=256: the heuristic's knob set fails NVRTC compilation (hidden by the frontend override)

  area:global_attentioncat-bugmod-backendorig-nv-eng
  NVIDIA/cudnn-frontend#1612 · Filter by author YangXu1990uiuc opened on Oct 13, 2026

• SDPA rng_dump at s_q == 1 leaves 75% of the dropout mask unwritten on SM80/SM89/SM90

  area:global_attentioncat-bugmod-backendorig-nv-eng
  NVIDIA/cudnn-frontend#1611 · Filter by author YangXu1990uiuc opened on Oct 13, 2026
  1

• SDPA dropout on SM100 cannot go below 1/256 and is biased high for small p

  area:global_attentioncat-bugmod-backendorig-nv-eng
  NVIDIA/cudnn-frontend#1610 · Filter by author YangXu1990uiuc opened on Oct 13, 2026

• SDPA: b or h > 65535 passes check_support/build_plans on every plan, then fails at kernel launch (native and FROST)

  NVIDIA/cudnn-frontend#1573 · Filter by author YangXu1990uiuc opened on Oct 11, 2026
  1

• sdpa fwd: execute-time override_strides are accepted but ignored (kernel uses the declared strides; silently wrong, cuDNN 9.27)

  area:global_attentioncat-bug
  NVIDIA/cudnn-frontend#1570 · Filter by author YangXu1990uiuc opened on Oct 11, 2026

• sdpa fwd: execute-time s_kv override off the KV tile attends to the tile's padding rows (silently wrong; first-ranked plans on H100/B200, cuDNN 9.27)

  area:global_attentioncat-bug
  NVIDIA/cudnn-frontend#1567 · Filter by author YangXu1990uiuc opened on Oct 11, 2026
  1

• SDPA: SM107 forward test387 watchdog after graph execution (padded D144/V104)

  NVIDIA/cudnn-frontend#1538 · Filter by author YangXu1990uiuc opened on Oct 9, 2026
  3

• test_mhas_v2 on cc 10.7 outgrew the SM107-test_mhas job's 2400 s timeout (13252 -> 17007 items)

  area:frostcat-ci
  NVIDIA/cudnn-frontend#1526 · Filter by author YangXu1990uiuc opened on Oct 9, 2026 · Frontend 1.33.0

• [FROST][SM107] BF16 paged sink decode remains 2.6–9x slower than vLLM FlashInfer GEN on identical page16 inputs

  NVIDIA/cudnn-frontend#1517 · Filter by author vedaanta opened on Oct 9, 2026

• [FROST][SM107] Selected BF16 sink SDPA plan is 1.5x slower than native backend for high-batch verification

  NVIDIA/cudnn-frontend#1472 · Filter by author vedaanta opened on Oct 8, 2026

• sdpa fwd (frost): nonpaged THD split rule overcounts CTAs (no GQA packing, ragged max length, one-wave cap); 1.4-3x left unselected

  NVIDIA/cudnn-frontend#1439 · Filter by author YangXu1990uiuc opened on Oct 7, 2026
  6

• frost(sdpa): forward kernels fold attn_scale into exp2 after an unscaled -inf-masked max, so attn_scale <= 0 with any mask gives NaN (SM80/SM100/SM107/SM120)

  NVIDIA/cudnn-frontend#1435 · Filter by author YangXu1990uiuc opened on Oct 7, 2026
  1